Hi Ard,
ilias.apalodimas@linaro.org wrote:
Hi Samer,
Thanks for having a look at this
On Mon, Mar 01, 2021 at 08:44:00PM +0000, Samer El-Haj-Mahmoud wrote:
-----Original Message----- From: Heinrich Schuchardt xypron.glpk@gmx.de Sent: Monday, March 1, 2021 2:39 PM To: Ilias Apalodimas ilias.apalodimas@linaro.org; Grant Likely Grant.Likely@arm.com Cc: Boot Architecture Mailman List boot-architecture@lists.linaro.org; Samer El-Haj-Mahmoud Samer.El-Haj-Mahmoud@arm.com; Ard Biesheuvel ardb@kernel.org; Leif Lindholm leif@nuviainc.com Subject: Re: EFI_LOAD_FILE2 for initrd standardization
...
The UEFI spec knows two types of separators for device paths. Both have type 0x7F (End of Hardware Device Path) but differ by the sub-type:
Sub-Type 0xff – End Entire Device Path Sub-Type 0x01 – End Instance of a Device Path
Field EFI_LOAD_OPTION.FilePathList[] is described in the UEFI spec as follows:
"A packed array of UEFI device paths. The first element of the array is a device path that describes the device and location of the Image for this load option."
It is not immediately clear if the separators between the array elements are of sub-type 0xff or 0x01. The description in the UEFI spec should be reworked for more clarity.
Agree that this is not clear, and could be interpreted either way. And yes, agree the UEFI spec needs a clarification
The current EDK II coding requires that the device path identifying the UEFI binary (i.e. FilePathList[0] is terminated by a sub-type 0xff end node.
The EDK2 code seems to be incomplete, with a "TODO" to support the FilePathList[]. In fact, the code calls it "FilePath" to be clear that it is assuming a single DevicePath (which means a 0xFF sub-type termination) https://github.com/tianocore/edk2/blob/master/MdeModulePkg%2FLibrary%2FUefiB...
Searching the code further, I see there is support for parsing multi-instance device path (separated by END_INSTANCE_DEVICE_PATH_SUBTYPE, or 0x1) in things like UefiDevicePathLib , parsing code, etc... But it does not seem to be supported at all in the UEFI Boot Manager implemented in EDK2
The UEFI variable ConDev is decribed as "The device path of all possible console input devices". The spec does not refer to it as an array of device paths.
Looking at EDK2, ConXDev are treated as multi-instance devices, with ConPlatformDxe using AppendDevicePathInstance() and GetNextDevicePathInstance() to construct / navigate the multiple-instances
The spec should have defined these as a "multi-instance device path of all possible console X devices". See my notes below on multi-instance DP.
So it seems that the UEFI spec editors mean by array of device paths that an element of the array is separated by sub-type 0xff. Each individual array element may be a collection of device paths instances separated by 0x01 sub-type end nodes.
I am leaning towards this conclusion as well. This is further supported by the following evidence:
Yes that's exactly what I thought as well and that's what option (2) was trying to describe on my first mail.
What I had in there was: Loaded Image device path - end node - VenMedia - Initrd DP - end instance - (repeat) - Initrd DP - end node - other DPs
- EFI_DEVICE_PATH_UTILITIES_PROTOCOL has a function called IsDevicePathMultiInstance() along with walker functions AppendDevicePathInstance(), GetNextDevicePathInstance(). The description of these functions make it clear that the intention is to treat 0x01 as a separator between multiple instances of device paths in a multi-instance device path structure, with 0xff as the final end of DP indicator. For example, AppendDevicePathInstance() is defined as:
" This function creates a new device path by appending a copy of the specified device path instance to a copy of the specified device path in an allocated buffer. The end-of-device-path device node is moved after the end of the appended device node and a new end-of-device-path-instance node is inserted between."
- In EFI_USER_INFO , there are user info policy types that leverage the device path definitions (such as EFI_USER_INFO_ACCESS_FORBID_LOAD), and they are clear on distinguishing the term "multi-instance device path" from a "series of UEFI device paths":
"The record is a series of normal UEFI device paths (not multi-instance device paths)."
- The "multi-instance" usage also shows up in the language of LocateDevicePath()
So if the spec intended for FilePathList[] to use the 0x1 separator, they should have used the term "multi-instance" device path, which is not the case.
In the device path spec a sentence could be added as follows:
"A packed array of UEFI device paths. The first element of the array is a device path that describes the device and location of the Image for this load option. *Each array element is terminated by a sub-type 0xff, End Entire Device Path node.*"
Yes. Or could simply say: "A packed array of UEFI device paths (not multi-instance device paths). The first element of the array is a device path that describes the device and location of the Image for this load option."
Either way, we can do a "code first" ECR in the UEFI spec. I say code first to ensure that it remains in public domain, and not blocked by UEFI Forum NDA until the publication of the next version of the UEFI spec. If we agree, I can get this process started and propose the language.
That's fine by me. I have most of the u-boot code ready, I'll finish it up and send patches.
Adding initrd or device tree device paths could be implemented as follows:
Array element [0]: device path of the binary (e.g. the Linux kernel) terminated by 0xff.
Array element [i], i > 0: device paths of the different initial RAM disks separated by 0x01 instance end nodes and terminated by an 0xff entire path node.
To identify the device paths with all its instances as initrds we can prepend a VenMedia() device path node with a specific GUID to the entire path.
Array element [j], j > 0: device path of the device-tree possibly followed by instances of device paths of device-tree overlays separated by 0x01 instance end nodes and terminated by an 0xff entire path node.
To identify the device paths with all its instances as device trees and device tree overlays we can prepend a VenMedia() device path node with a specific GUID to the entire path.
One suggestion here is to take into consideration: whether we standardize this proposal (for loading additional kernel files from element[i], i>0 ) or not, the FW implementation must still be compliant with the UEFI rules around booting from a "short form device path" (UEFI spec, section 3.1.2)
Ok that's something I'll have to change since I am inserting the entire DP, but as Heinrich pointed out, u-boot already has support for short form device paths so the change shouldn't be too hard.
If I am understanding this correctly, the intent is to use the first devicepath in the Boot#### variable as the boot image, as is the case today, and use subsequent pairs of devicepaths as a key/value store, where the key is an abstract device path which describes how uboot should expose the actual device path in the value, using LoadFile2 and the existing Linux initrd vendor GUID + an filepath specifier.
You do understand it correctly, there's slight variation on what we can do to store those device paths on the FilePathList[] but that's more an implementation detail. Both would still have the same result and it only affects how the bootloader will parse those (and possibly concat multiple initrds) before copying them to the kernel provided buffer.
I am fine with all this as long as there are no changes to the way the kernel consumes the LoadFile2 protocol. The kernel does not reason about Boot#### variables in the first places, and so how uboot uses them to keep track of boot entries is of no relevance to the loader.
I think repurposing LoadFile2 for other OSes or hypervisors like this makes sense, and there was already some pull from the Xen folks to implement GRUB multiboot in this manner. The only hard requirement I have from the Linux side is that loading LoadFile2+end_node yields the Linux initrd as it does today. What it does for LoadFile2+whatever (including where whatever == /initrd) does not impact Linux so it is free to do whatever is needed for the other use cases you have in mind.
I think what's hard for people to understand here, and part of the confusion comes from the part that the efi-stub will eventually call the protocol with the end node. So the discussions I've been having around is "why does the kernel call that with and end node and not the initrd GUID" (by initrd GUID I mean the "struct efi_vendor_dev_path vendor" defined in the kernel).
But what we need to make clear here, is that this device path is *already used* to locate the protocol. After that and following the EFI spec, the 'front' part is removed and the remaining portion is returned. So the returned device path (which is now just the end node) is used by the kernel to call that protocol. This identifies the 'initrd GUID uniquely' and a similar method can be used for other files.
Hope that makes sense now
Regards /Ilias
If