On Mon, Nov 1, 2021 at 12:11 AM Andrew Jones <drjones@xxxxxxxxxx> wrote: > > On Sun, Oct 31, 2021 at 02:54:00PM -0700, Zixuan Wang wrote: > > On Sun, Oct 31, 2021 at 9:14 AM Marc Orr <marcorr@xxxxxxxxxx> wrote: > > > > > > On Sun, Oct 31, 2021 at 12:28 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > > > > > > > On 31/10/21 06:56, Zixuan Wang wrote: > > > > > Hello, > > > > > > > > > > This patch series refactors the x86_64 UEFI set up process and fixes the > > > > > `run-tests.sh` script to run under UEFI. The patches are organized as > > > > > three parts. > > > > > > > > > > The first part (patches 1-2) refactors the x86_64 UEFI set up process. > > > > > The previous UEFI setup calls arch-specific setup functions twice and > > > > > generates arch-specific data structure. As Andrew suggested [1], we > > > > > refactor this process to make only one call to the arch-specific > > > > > function and generate arch-neutral data structures. This simplifies the > > > > > set up process and makes it easier to develop UEFI support for other > > > > > architectures. > > > > > > > > > > The second part (patch 3) converts several x86 test cases to > > > > > Position-Independent Code (PIC) to run under UEFI. This patch is ported > > > > > from the initial UEFI support patchset [2] with fixes to the 32-bit > > > > > compilation. > > > > > > > > > > The third part (patches 4-7) fixes the UEFI runner scripts. Patch 4 sets > > > > > UEFI OVMF image as readonly. Patch 5 fixes test cases' return code under > > > > > UEFI, enabling Patch 6-7 to fix the `run-tests.sh` script under UEFI. > > > > > > > > > > This patch set is based on the `uefi` branch. > > > > > > > > Thank you, for patches 1-6 I have squashed the patches when applicable > > > > (1, 4, 5, 6) and queued the others (2 and 3). > > > > > > > > I did not queue patch 7 yet, it seems okay but I want to understand > > > > better the changes it needs in the harness and what is missing. I'll > > > > take a look during the week. > > > > > > SGTM, thank you! Zixuan and I discussed a few things that are missing: > > > > > > 1. Test cases that take the `-append` arg are currently marked `SKIP`. > > > Two issues need to be resolved here. First, we're not using QEMU's > > > `-kernel` flag for EFI test cases [1]. And the `-append` flag does not > > > work without the `-kernel` flag. I don't understand the details on why > > > we don't use the `-kernel` flag myself. Maybe Zixuan can elaborate. > > > Second, assuming we fix the first issue, then we need to enlighten the > > > KVM-Unit-Tests under UEFI to parse kernel command line arguments and > > > pass them down to the test cases via `argv`. Zixuan pointed out to me > > > that there is some prior work from Drew [2] that we should be able to > > > follow to make this work. So I'm hoping that Zixuan and I can work > > > together on solving these issues to get the argument passing working > > > next. > > > > Thank you for the detailed summary! > > > > Current kvm-unit-tests pass an EFI binary as part of a disk image, > > instead of using the `-kernel` argument. > > > > I just tested the `-kernel` argument and it seems to work with EFI > > binaries, and more importantly, it's really fast (bypassing the > > default 5-second user input waiting). I will update the `x86/efi/run` > > to use `-kernel` argument to pass the EFI binaries. > > > > Since `-kernel` is working, I can start to investigate how to use > > `-append` to pass arguments. If that doesn't work well, an alternative > > approach could be: > > > > 1. (host) create a file `args.txt` in the disk image, which contains > > all the arguments needed > > 2. (guest) call UEFI filesystem interface to read this `args.txt` from > > the disk image, parse it and pass the arguments to `main()` > > > > > 2. We need a way to annotate test cases in `x86/unittests.cfg` as > > > known to work under SEV. I'm thinking of doing this via new (very > > > broad) test groups in `unittests.cfg`. I _think_ SEV is the primary > > > scenario we care about. However, folks may care about running the test > > > cases under UEFI outside of SEV. For example, last time I checked, > > > emulator runs OK under UEFI minus SEV-ES but fails under SEV-ES. And > > > similarly, while most test cases work under UEFI minus SEV, there are > > > a few that do mis-behave -- and it probably makes sense to document > > > this (e.g., via annotations in `unittests.cfg`). Also, there are many > > > variations of SEV (SEV, SEV-ES, SEV-SNP)... And hopefully some of this > > > will eventually be applicable to TDX as well. So many testgroups is > > > not a good solution. I'm not sure. > > > > Adding an `efi` group seems helpful. E.g., the current `x86/smap.c` > > does not work under UEFI; but the `run-tests.sh` still tries to run > > this test case, even if this test case is not compiled. > > > > > 3. Multi-CPU needs to be made to work under UEFI. For now, patch #7 > > > forces all EFI test cases to run with 1 vCPU. I chatted with Brijesh, > > > and he mentioned that Varad would like to work on this. However, if > > > anything here changes, please let me know, because we can work on this > > > as well. But for now, I'm not planning to work on it so we can avoid > > > duplicating work. > > > 4. UEFI runs a lot slower than SEABIOS. It doesn't help that the test > > > harness launches QEMU more than once for each test case (i.e., it runs > > > the `_NO_FILE_4Uhere_` scenario to check QEMU arguments). I'm not sure > > > how much of an issue this is in practice. Depending on the answer, I > > > know Zixuan had some ideas on how to speed this up in the current test > > > harness. Or maybe we can explore an alternative to the > > > `_NO_FILE_4Uhere_` approach instead. > > > > As the `-kernel` argument now works with the EFI binaries and is > > significantly faster, this should not be an issue anymore. We just > > need to update the runner scripts to use `-kernel` argument. > > You can add an additional '-kernel' + EFI binary runner if you want, but > the goal of being able to run kvm-unit-tests on bare-metal means we > shouldn't be counting on QEMU/OVMF to do magic stuff with the kernel. We > need to build disk images. Argument passing works with EFI apps, when > implemented, so that's not a problem. I also created a script that uses > the framework's for_each_unittest to generate an EFI script that allowed > each test to be easily run with its arguments. > > Thanks, > drew I see, I think an alternative approach is to rename test case binaries to UEFI default binary filename, which is EFI/BOOT/BOOTX64.EFI for x86_64. This should work just like the `-kernel` argument. I will explore this approach with the argument passing mechanisms. Best regards, Zixuan
