[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Minios-devel] Some considerations of ARM Unikraft supports





On 07.02.2018 07:16, Wei Chen wrote:
Hi Simon,

-----Original Message-----
From: Simon Kuenzer [mailto:simon.kuenzer@xxxxxxxxx]
Sent: 2018年2月7日 0:34
To: Wei Chen <Wei.Chen@xxxxxxx>; Julien Grall <julien.grall@xxxxxxxxxx>
Cc: Felipe Huici <Felipe.Huici@xxxxxxxxx>; Kaly Xin <Kaly.Xin@xxxxxxx>; Shijie
Huang <Shijie.Huang@xxxxxxx>; Florian Schmidt <Florian.Schmidt@xxxxxxxxx>;
Costin Lupu <costin.lup@xxxxxxxxx>; nd <nd@xxxxxxx>; minios-
devel@xxxxxxxxxxxxx
Subject: Re: [Minios-devel] Some considerations of ARM Unikraft supports

Hi Wei,

On 06.02.2018 08:58, Wei Chen wrote:
Hi Simon,

-----Original Message-----
From: Simon Kuenzer [mailto:simon.kuenzer@xxxxxxxxx]
Sent: 2018年2月6日 0:21
To: Wei Chen <Wei.Chen@xxxxxxx>; Julien Grall <julien.grall@xxxxxxxxxx>
Cc: Felipe Huici <Felipe.Huici@xxxxxxxxx>; Kaly Xin <Kaly.Xin@xxxxxxx>;
Shijie
Huang <Shijie.Huang@xxxxxxx>; Florian Schmidt <Florian.Schmidt@xxxxxxxxx>;
Costin Lupu <costin.lup@xxxxxxxxx>; nd <nd@xxxxxxx>; minios-
devel@xxxxxxxxxxxxx
Subject: Re: [Minios-devel] Some considerations of ARM Unikraft supports

Hi Wei, hi Julien,

thanks a lot for discussing this already, I put my comments inline.

On 05.02.2018 08:22, Wei Chen wrote:
Hi Julien,

Thanks for your comments!
Replies inline.

-----Original Message-----
From: Julien Grall [mailto:julien.grall@xxxxxxxxxx]
Sent: 2018年2月2日 18:43
To: Wei Chen <Wei.Chen@xxxxxxx>; Simon Kuenzer <simon.kuenzer@xxxxxxxxx>
Cc: Felipe Huici <Felipe.Huici@xxxxxxxxx>; Kaly Xin <Kaly.Xin@xxxxxxx>;
Shijie
Huang <Shijie.Huang@xxxxxxx>; Florian Schmidt <Florian.Schmidt@xxxxxxxxx>;
Costin Lupu <costin.lup@xxxxxxxxx>; nd <nd@xxxxxxx>; minios-
devel@xxxxxxxxxxxxx
Subject: Re: [Minios-devel] Some considerations of ARM Unikraft supports

Hi,

On 02/02/18 09:10, Wei Chen wrote:
This week I am trying to boot Unikraft on ARM64/KVM platform. In this
progress I have
got some considerations and written a simple proposal:

My first target is to enable Unikraft on ARM64+Kvm, so this proposal
would
focus on ARM64+Kvm.
But the goal of ARM support is to enable Unikraft on ARM32/ARM64 based
hypervisors (ARM32/64 Kvm,
ARM64 Xen and etc). So we have to consider to keep current multi-arch
framework and reuse common
code like virtual drivers for ARM32/ARM64.

1. Modify the folders for multi-architectures
        1.1. Add arm64 folder to unikraft/arch:
             unikraft----arch----arm
                           |-----x86_64
                           |-----arm64 <-- New

             Above folders contains architecture specified Makefile,
Config,
Compiler flags and some
             code. In most cases, these files are exclusive. So we'd
better
keep each arcitecture in
             a standalone floder. This also can avoid doing to much
changes
to
Unikraft Makefile.

             If we add arm64 to unikraft/arch/arm, we have to do more
ARCH
comparasion in Makefile:
             unikraft----arch----arm----arm32
                               |      |-----arm64 <-- New
                         |
                       |-----x86_64
             Before:$(UK_BASE)/arch/$(ARCH)/Makefile.uk.
             After:$(UK_BASE)/arch/arm/$(ARCH)/Makefile.uk
             This change is complex, so we'd better to add arm64 folder
to
unikraft/arch.

Except the assembly code, most of the C code should be very similar
between ARM64 and ARM32. So it might make more sense to have a directory
arch/arm with sub-folder arm32 and arm64.


This is one option I had considered. But this will add a new variable
(VENDOR) to
make scripts. e.g. :$(UK_BASE)/arch/$(VENDOR)/$(ARCH)/Makefile.uk
And currently, only architecture dependent code will be placed in $(ARCH)
folder.
For example, in arm folder, there are some files for arm32 math library.
These
files can only be used for arm32.

What is this vendor variable about? Is it something that applies to a
specific silicon? Is it required to add subfolders for it?


Yes, it applies to a specific silicon. But "VENDOR" is not very accurate
here.
I had considered it again, because x86 is not a "VENDOR", and not all x86
chips
Belong to intel, Maybe use "FAMILY" is better.

If we really have some common C code for ARM32/64, I agree to add subfolders
for it.

unikraft----arch----arm----arm32  ARM family arm32 and arm64 architectures
               |       |-----arm64
               |
               |------x86----i386
                       |-----x86_64 X86 family i386 and x86_64 architectures


Sorry, I forgot to mention that you also should add only code here which:
1) ...is exposed to the user with an interface in include/uk/arch/*
2) ...works with all platforms (including linuxu which is special).
     So for instance, you should not add code that uses privileged
     instruction that could not be executed in Linux userspace. If there
     is a different implementation needed, it is a hint that this
     functionality need to be moved to the platform API
     (include/uk/plat/*)


Ahh, I understand now. Thanks for your explanation.

I had a discussion with Costin, and we were thinking of placing code
that is shared by multiple platforms (but not by all, or is not
architecture code) in plat/common/arm/* and plat/common/arm/arm64/*.
Your platforms libs would include the source files from this directory.

Subdirectories (for e.g., timer, GIC) are fine. What do you think? If
you agree we will put a commit that introduces a structure to the
staging branch.


I think this idea is good. But the example here is not very accurate ; )
Once the "drivers" folder has been introduced, I still want to move the
timer, GIC to it.


Hum. You are right, we should probably distinguish which drivers go bundled to the platform libraries and which drivers are a selectable option and stay as independent library. This is not clear at all yet.

What would you guys think if we do the following:

plat/common/arm/* <-- code that is shared among multiple ARM platform
                      libs (probably includes bare essential drivers
                      like interrupt controllers and timers for
                      scheduling)
plat/common/x86/* <-- same for x86 platform libs
plat/common/drivers/* <-- device and bus drivers that are going to be
                          built as individual libraries
                          (e.g., NIC, block device drivers)
plat/common/drivers/include/* <-- Include folder for driver APIs that
                                  depend on each other (for example:
                                  PCI bus so that e1000 works with
                                  pcifront but also linuxu's VFIO-based
                                  pci bus)

Note that the NET or BLOCK device API (that are implemented by individual drivers) should be defined by libraries in libs/ (e.g., lib/uknet, lib/ukblock; network stacks would then use uknet for doing networking I/O, VFSs would use ukblock).

The structure of the drivers folder is still not clear though. How should we organize the sub structure? Would maybe something similar to Linux's drivers folder make sense? I think people might be most familiar with this.

If we have this, each of the platform Config.uk's would list only a subset of drivers that they can work with (e.g., pcifront on the Xen platform lib only). We also have to figure out how we handle Makefile.uk's and Config.uk's for a driver library. Probably we need global switches for each driver that can enable by one or multiple platforms. A new menu item (either in the root or platform structure) should appear that lists only enabled drivers and allows us to configure each of them individually. The platform's Linker.uk would then need to include the depending and compiled driver library objects to the final linking.

@Wei, Costin: What do you think? Does this makes sense to you?
I think the best way to go with this question The best might be to go just with this and see if it fits our needs. If not, we restructure it afterwards.


If some C codes are very similar between arm32 and arm64, I think this
code
would
be very similar between arm and x86 too. We can place these codes in
Unikraft/lib.

Above 2 options would affect the common framework, so I still want to get
some
Comments from Simon.

I welcome this discussion because one of the exercises of Unikraft's 0.2
releases is to figure out how to do the right split.
I am okay with changing the structure of the arch folder substructure if
we can foresee already that it will make sense. In such a case, I would
also like to adopt the same principle to the x86 architecture folder.

The idea of architecture libraries is that they contain code which is
only special to the CPU but the same to all of the target platforms
(xen, kvm, linux). We were originally expecting that this is mostly
assembly code but we might be wrong with our original assumption. So, if
you foresee any common C code for 32 and 64bit ARM that would be
duplicated otherwise, we should use a single arm folder instead.


Sorry, about " use a single arm folder instead". Does it mean we don't add
Any subfolders to arm or x86 folder? Like following?

unikraft----arch----arm
               |
               |------x86


Sorry, I wasn't clear. I meant:
arch/arm/*

with specific code in:

arch/arm/arm32
arch/arm/arm64


Thanks for your clarification, I got it now.



        1.2. Add arm64 to unikraft/include/uk/arch

        1.3. Add arm64 kvm platform code to unikraft/plat/kvm/arm, and
use
Makefile to select
             objects for correct architecutre:

             ifeq ($(ARCH_X86_64),y)
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/x86/entry64.S
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/x86/cpu_x86_64.c
             else ifeq ($(ARCH_ARM_64),y)
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/arm/entry64.S
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/arm/cpu_arm64.c
             else ifeq ($(ARCH_ARM_64),y)
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/arm/entry.S
                LIBKVMPLAT_SRCS-y += $(LIBKVMPLAT_BASE)/arm/cpu_arm.c
             endif

        1.4. Add a "drivers" folder to unikraft/
             This because we may have some virtual device drivers can be
shared
among platforms.
             For example, we can reuse virtual uart, timer and gic
drivers
from
arm32/arm64 Kvm/xen.

Is it okay for you to wait with the driver folder a bit? I am currently
working on PCI for x86 KVM and I figured that Unikraft need an mechanism
to select drivers for devices (and maybe buses) individually for each
platform. But drivers are still something that depend on the platform.
For instance Xen could reuse the same PCI drivers with pcifront, linux
with VFIO, but a third platform might not support PCI at all.

Because of this, I am currently considering to introduce an folder in
plat: e.g., plat/common/drivers/pci/virtio-net. What do you guys think?


That's quite good, I will wait it : )


2. Bootloader
        2.1. Because of the BIOS, x86 is using multiboot to load kernel
on
Linux-KVM QEMU. But on ARM platforms,
             we can skip the EFI and boot from the Virtual Machine's RAM
base
address. So we can place _libkvmplat_entry
             to the CPU's reset entry by link script. On ARM64 platform,
the
default virtual machine CPU model is cortex A15.

Cortex A15 does not support 64-bit. So how come it is the default
virtual machine CPU model for ARM64?


   From the code, if we don't specify any cpumodel, the mach-virt's default
cpumodel will be set to "cortex-a15". But you'are right, if we use cortex-
15
by default, we can run any 64-bit image. Here is my mistake. We have to
set
correct cpumodel (cortex-a53/a57 or host) in command line to make 64-bit
image
work. But the mach-virt is still using the a15memmap and a15irqmap.


But likely, you want to expose the same MIDR as the underlying CPU. So
if an errata has to be implemented in Unikraft, it will be able to know
it.


Exposing the underlying CPU's MIDR to guest is depending on the
hypervisors.
For Unikraft itself, it doesn't know whether this MIDR is the same as the
underlying
CPU or not. And actually, no matter what cpumodel the hypervisor is
emulating, the
code is running on the physical CPU directly. We don't emulate the CPU
instructions.
If we run Unikraft on a corext-a53 host CPU, we can compile this image
with
gcc flags
like fix-a53-error.


             plat/kvm/arm/link64.ld:
             ENTRY(_libkvmplat_entry)
             SECTIONS {
                 . = 0x40000000;

                 /* Code */
                 _stext = .;

                 .text :
                 {
                     *(.text)
                     *(.text.*)
                 }

                 _etext = .;
                 ...
             }

        2.2. Use the fixed physical addresses of PL011 uart, timer and
GIC.
So
we can skip the device tree parse.

What does promise you the PL011, timer, GIC will always be at the same
address?

My original idea was that we selected a fixed machine (mach-virt) for
Unikraft to run.
In this case, the memory map is fixed.

Or do you expect the user to hack unikraft build system to set
the address?


For my opinion, Yes. Why should we need to parse the device tree to
increase
our boot
time and footprint?


Sorry for my stupid question: Would this hardcode the guest device
configuration that you would need to set with KVM? I mean, how are
network devices (or other) are handover to the guest? If yes, I am
concerned that Unikraft is getting difficult to use on ARM. I would
rather prefer to provide a configuration option where users could
disable that the image scans the device tree and expects devices at
hardcoded places.

While I was writing this proposal, I hadn't consider so many devices. I just
considered some platform devices like interrupt controller, timer and UART.
At that moment, I prefer to hardcode. But now I think parse the device tree
is better. Because the virtual net/block devices are dynamic configuration
for a VM.


Good. Unikraft has libfdt already included. You probably should use this
one for doing the parsing and depend the platform libraries on it (see
arm32 platforms).


At least from Xen PoV, the memory layout is not part of the ABI and a
guest should rely on the DT for getting the correct addresses.


I understand your concern. It's not a part of the ABI. So the addresses
can
be changed
for different boards.

I think we must do a tradeoff between flexibility and deploy density (boot
time and footprint)


If this makes sense for you: I prefer having the most flexible as
default and provide configuration options with Config.uk to switch them
off individually. I think Unikraft should handover such tradeoff
question to Unikernel builders.


That would be good.


Perfect ;-)

        2.3. Setup exception traps.

3. Support single CPU.

This is fine for the first version. The other platforms also just
support a single CPU for now.


4. Support multiple threads.
        4.1. Implement GIC interrupt controller drivers. If we doesn't
specify
the gic version in QEMU's parameter,
             default GIC will be detected by kvm_arm_vgic_probe. Most ARM
hosts
are using GICv2, GICv3 and GICv4,
             and QEMU will provide GICv2 and GICv3 emulators. For best
compatibility, we have to implement gicv2
             and gicv3 drivers without MSI/MSI-X support. This means we
don't
need to implement gicv2m, gicv3-its
             for Unikraft in this time.
        4.2. Implment ARMv8 virtual timer driver.


Please contact Costin what is required from the Unikraft's scheduler
API. I CC'ed him.


Thanks, I will contact Costin when I start to implement this driver.

5. Setup a 1:1 mapping pagetable for Physical memory and Virtual memory.
        5.1. Configure MMU
        5.2. Create page tables with 1GB or 2MB block


Good.

6. Implement PSCI interface to support machine shutdown.

FWIW, system_off only exist from PSCI 0.2 and onwards.


It seem the psci-0.2 is the default PSCI version of mach-virt with KVM.


7. Network, block and etc IO devices?
       Should we have to port virtual device driver like virtio-net, pv-
net
from KVM and Xen?

After we agreed how Unikraft should include drivers we can start with
porting them. Is KVM on ARM using virtio-net, too? Is there a virtual
PCI bus attached?

Yes, KVM on ARM is using virtio-net too. The virtio-net is connect to a
virtio-mmio bus. But there is a ECAM PCI host controller emulator too.


How are other devices attached? For instance block devices. I remember
we have SD card emulation. Maybe we need another bus driver that uses
FDT later to make them work in Unikraft.


By default, all virtio devices will attach to virtio-mmio bus. PCI Pass through
devices can be connected to ECAM PCI host emulate. So if we want to support
ARM PCI pass through, we have to implement ECAM PCI host driver for Unikraft.

If you want to add a SD Card controller to VM. This controller may attach to
Platform bus or simple-bus.
         SD_MMC_1@B000000 {         ===>> attach SD MMC to platform bus
                compatible = "SD1...";
         }
platform@c000000 {
                 compatible = "qemu,platform", "simple-bus";
                 ranges = <0x0 0x0 0xc000000 0x2000000>;
                 interrupt-parent = <0x8001>;
                 #address-cells = <0x1>;
                 #size-cells = <0x1>;

                 SD_MMC_2@c003000 { ===>> attach SD MMC to simple bus
                        compatible = "SD2...";
                 }
         };

Both of above buses are very simple. We should implement them for Unikraft.
But I am not sure what is the "SD card emulation" meaning? Is it a SD card
Controller emulator for guest or just a block device? If it's a block device,
Why should we have to care about is it a SD card or not?



Hey, thanks for the clarification. For you question: Maybe I used the wrong words. I meant this SD card reader entries in dtb that are used for attaching block devices to the guest - and emulated by QEMU. Is this way of attaching block devices the default way for ARM?



There are no emulation provided on Xen, so you would need PV drivers to
get access to the network/block.

This is fine ;-).


Yes, I have the same opinion with you 😊



Cheers,

--
Julien Grall

Thanks,

Simon

Thanks,

Simon

_______________________________________________
Minios-devel mailing list
Minios-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/minios-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.