RFE: add RISC-V 32-bit arch support

[kraj]

Update syscall.csv to reflect rv32 arch and use latest 5.13-rc5 kernel for that

[coveralls]

Coverage increased (+0.01%) to 89.615% when pulling 0637247 on kraj:riscv32 into a792221 on seccomp:main.

[drakenclimber]

I think the changes look really good. Thanks, @kraj.

It looks like we forgot to update this list in arch-syscall-validate when we added RISCV64. (Github won't let me comment on lines that haven't been modified.) We should update it with both RISCV64 and RISCV32.

I don't have access to a RISCV32 box, and obviously our continuous integration isn't setup to run it either. Do the following commands pass on RISCV32:

# ./configure --enable-python
# make check
# (cd tests; ./regression -T live)

[kraj]

I think the changes look really good. Thanks, @kraj.

It looks like we forgot to update this list in arch-syscall-validate when we added RISCV64. (Github won't let me comment on lines that haven't been modified.) We should update it with both RISCV64 and RISCV32.

I don't have access to a RISCV32 box, and obviously our continuous integration isn't setup to run it either. Do the following commands pass on RISCV32:

# ./configure --enable-python
# make check
# (cd tests; ./regression -T live)

Here is summary on qemu

Regression Test Summary
tests run: 4388
tests skipped: 149
tests passed: 4385
tests failed: 3
tests errored: 86

Failed tests
Test 55-basic-pfc_binary_tree%%001-00001 result: FAILURE 55-basic-pfc_binary_tree.sh rc=1

There is a batch of tests which errored with same message like

 batch name: 53-sim-binary_tree                                                                                                                                                             
 test mode:  c                                                                                                                                                                              
 test type:  bpf-sim                                                                                                                                                                        
 test arch:  x86_64                                                                                                                                                                         
Test 53-sim-binary_tree%%001-00001 result:   ERROR 53-sim-binary_tree rc=14                                                                                                                 
 test arch:  x86_64                                                                                                                                                                         
...

 batch name: 36-sim-ipc_syscalls                                                                                                                                                            
 test mode:  c                                                                                                                                                                              
 test type:  bpf-sim                                                                                                                                                                        
 test arch:  x86                                                                                                                                                                            
Test 36-sim-ipc_syscalls%%001-00001 result:   ERROR 36-sim-ipc_syscalls rc=14                                                                                                               
...

Does this look ok ? I am no way expert in libseccomp but if you give some hints on fixing these tests I can try
Let me know

[kraj]

I also ran tests on my pc and it came all passing

Regression Test Summary                                                                       
 tests run: 7443    
 tests skipped: 46                                                                            
 tests passed: 7443                                                                           
 tests failed: 0                     
 tests errored: 0                                                                             

[drakenclimber]

Here is summary on qemu

Thanks!

Failed tests
Test 55-basic-pfc_binary_tree%%001-00001 result: FAILURE 55-basic-pfc_binary_tree.sh rc=1

This failure doesn't surprise me. The syscall numbers won't align between RISC-V 32 and x86_64, so that's likely the culprit.

batch name: 53-sim-binary_tree
test mode: c
test type: bpf-sim
test arch: x86_64
Test 53-sim-binary_tree%%001-00001 result: ERROR 53-sim-binary_tree rc=14
test arch: x86_64

This failure is more interesting, but I would be fine with ignoring it for this patchset. (Adding a github issue might be a good idea, though). The first check of test 53 expects a return code of ERRNO(0) - which now that I think about it doesn't really make sense. Looks like RISC-V 32 got EFAULT instead.

batch name: 36-sim-ipc_syscalls
test mode: c
test type: bpf-sim
test arch: x86
Test 36-sim-ipc_syscalls%%001-00001 result: ERROR 36-sim-ipc_syscalls rc=14

This error could be the most concerning. Does RISC-V 32 multiplex its IPC and/or socket syscalls? See here and here for other architecture examples. You may need to look through the RISC-V kernel code to truly verify.

Does this look ok ? I am no way expert in libseccomp but if you give some hints on fixing these tests I can try

Yes, we're definitely getting closer.

Could you please look into the IPC issue? I'll try to create a RISC-V 32 qemu VM. I would prefer to resolve these issues if they aren't too difficult - famous last words :)

Thanks for your help!

[drakenclimber]

@kraj - I was able to run the automated libseccomp tests on a RISC-V 64-bit Fedora qemu image. They all passed :)

I half-heartedly looked around for a 32-bit RISC-V Linux image (from any distro) but didn't find one. What are you running?

[pcmoore]

One quick comment, please change the subject line on commit ee4aba3; something like "syscalls: update the syscall defs for Linux v5.13-rc5" or similar. We try very hard to stick to the "<subject_area>:<brief_description>" format.

Thanks.

[kraj]

One quick comment, please change the subject line on commit ee4aba3; something like "syscalls: update the syscall defs for Linux v5.13-rc5" or similar. We try very hard to stick to the "<subject_area>:<brief_description>" format.

Thanks.

done

[kraj]

@kraj - I was able to run the automated libseccomp tests on a RISC-V 64-bit Fedora qemu image. They all passed :)

I half-heartedly looked around for a 32-bit RISC-V Linux image (from any distro) but didn't find one. What are you running?

I am using yocto to build 32bit qemu image. I can upload one for you if you like

[drakenclimber]

I am using yocto to build 32bit qemu image. I can upload one for you if you like

@kraj - That would be awesome. Thanks!

[kraj]

I am using yocto to build 32bit qemu image. I can upload one for you if you like

@kraj - That would be awesome. Thanks!

@drakenclimber I have uploaded needed artifacts into https://uclibc.org/~kraj/qemuriscv32/
there is run.sh script which can help in booting it in qemu. Let me know how it goes.

[drakenclimber]

I'm not sure what's going wrong with my RISC-V 32 qemu setup. (RISC-V 64 worked fine for me.) It's hanging between the hardware output and the first dmesg line from the kernel. I've tried rebuilding a couple different versions of qemu and a couple different kernels.

I'll try and revisit it next week sometime.

[kraj]

I'm not sure what's going wrong with my RISC-V 32 qemu setup. (RISC-V 64 worked fine for me.) It's hanging between the hardware output and the first dmesg line from the kernel. I've tried rebuilding a couple different versions of qemu and a couple different kernels.

I'll try and revisit it next week sometime.

I build qemu also using yocto, but this should have worked with system qemu too from your distro ( fedora/debian) If needed
perhaps you can build yocto I can help.

[drakenclimber]

Getting closer. I switched over to using your kernel. (In fact I'm using all of the files you provided verbatim.)

Now it's puking on my rootfs:

[    1.551081] EXT4-fs (vda): bad geometry: block count 725665 exceeds size of device (695776 blocks)
[    1.554322] EXT4-fs (vda): bad geometry: block count 725665 exceeds size of device (695776 blocks)
...
[    1.604849]  driver: virtio_blk
[    1.605066] No filesystem could mount root, tried: 
[    1.605081]  ext3
[    1.605223]  ext2
[    1.605282]  ext4
[    1.605339]  btrfs
[    1.605399] 
[    1.605765] VFS: Unable to mount root fs on unknown-block(253,0)
[    1.606136] User configuration error - no valid root filesystem found

[drakenclimber]

Getting closer. I switched over to using your kernel. (In fact I'm using all of the files you provided verbatim.)

Now it's puking on my rootfs:

[    1.551081] EXT4-fs (vda): bad geometry: block count 725665 exceeds size of device (695776 blocks)
[    1.554322] EXT4-fs (vda): bad geometry: block count 725665 exceeds size of device (695776 blocks)
...
[    1.604849]  driver: virtio_blk
[    1.605066] No filesystem could mount root, tried: 
[    1.605081]  ext3
[    1.605223]  ext2
[    1.605282]  ext4
[    1.605339]  btrfs
[    1.605399] 
[    1.605765] VFS: Unable to mount root fs on unknown-block(253,0)
[    1.606136] User configuration error - no valid root filesystem found

Got it booting after a resize2fs :)

$ resize2fs -MPf yoe-sdk-image-qemuriscv32-20210812183746.rootfs.ext4
resize2fs 1.45.6 (20-Mar-2020)
Estimated minimum size of the filesystem: 725665

[drakenclimber]

Thanks for all the help, @kraj. I have my RISC-V 32 qemu image up and running. I found a couple issues with the patchset, but there's still more work to do. (I need to switch gears and work on a couple issues for the v2.5.2 release of libseccomp.)

The arch_syscall_resolve_num() function is returning NULL for some RISC-V 32 syscalls. You need to add an entry in syscalls.csv in every row for RISC-V 32.

Here are a couple changes I made that helped the tests get further. YMMV. Let me know if you have any questions

diff --git a/include/seccomp-syscalls.h b/include/seccomp-syscalls.h
index 01554fe..12ecc4d 100644
--- a/include/seccomp-syscalls.h
+++ b/include/seccomp-syscalls.h
@@ -283,6 +283,7 @@
 #define __PNR_clock_nanosleep                  -10249
 #define __PNR_gettimeofday                     -10250
 #define __PNR_quotactl_path                    -10251
+#define __PNR_fcntl                            -10252


diff --git a/src/arch-riscv32.c b/src/arch-riscv32.c
index 53b3126..10418f4 100644
--- a/src/arch-riscv32.c
+++ b/src/arch-riscv32.c
@@ -24,8 +24,8 @@ const struct arch_def arch_def_riscv32 = {
        .token_bpf = AUDIT_ARCH_RISCV32,
        .size = ARCH_SIZE_32,
        .endian = ARCH_ENDIAN_LITTLE,
-       .syscall_resolve_name = riscv32_syscall_resolve_name,
-       .syscall_resolve_num = riscv32_syscall_resolve_num,
+       .syscall_resolve_name_raw = riscv32_syscall_resolve_name,
+       .syscall_resolve_num_raw = riscv32_syscall_resolve_num,
        .syscall_rewrite = NULL,
        .rule_add = NULL,
 };

[kanavin]

@kraj can you please take a look at the comments - 2.5.2 is now out, and the patches that are carried in Yocto for 2.5.1 are difficult to forward-port. So we can't update.

[kraj]

@kraj can you please take a look at the comments - 2.5.2 is now out, and the patches that are carried in Yocto for 2.5.1 are difficult to forward-port. So we can't update.

let me do the update for that package.

[drakenclimber]

Thanks, @kraj. I think you're making good progress. The tests passed on libseccomp's (x86_64) CI, and I ran them on an arm VM where they also passed. I encountered several errors on RISC-V 32, though:

Regression Test Summary
 tests run: 4370
 tests skipped: 149
 tests passed: 4367
 tests failed: 3
 tests errored: 108

The failures were largely in socket, IPC, and the binary tree logic. How does RISC-V 32 handle multiplexing of sockets/IPC?

Finally, combining the update to Linux v5.15.0-rc3 and the addition of RISC-V 32 into one commit may cause maintenance issues. If we bisect a problem down to that commit, is it because of the kernel change or the RISC change?

[kanavin]

@kraj can you please rebase and fix the conflicts again, so we can update to 2.5.3? Would be good to get this merged as well.

[kraj]

@drakenclimber rebased again. I will test again on riscv32 and try to dig into reasons for failures you saw.

[drakenclimber]

@drakenclimber rebased again. I will test again on riscv32 and try to dig into reasons for failures you saw.

Thanks, @kraj. I have my riscv32 VM up and running again. I'll check out the changes this week.

[drakenclimber]

I ran the automated tests on my riscv32 VM. (Out of expediency, I didn't enable python for this test run.)

Regression Test Summary
 tests run: 4353
 tests skipped: 149
 tests passed: 4350
 tests failed: 3
 tests errored: 118
============================================================
FAIL: regression
==================
1 of 1 test failed
==================

Looks like there were failures in the binary tree, IPC, and socket-related tests.

[pcmoore]

From what I can see this PR has some test failures, but no follow up with fixes; any updates @kraj?

[kraj]

From what I can see this PR has some test failures, but no follow up with fixes; any updates @kraj?

yeah I had to shift my focus on other things, I can revive this again perhaps next week or so.