Add support for CheriBSD #36
Conversation
The current version of epoll-shim does not work as a purecap library. Use my fork until jiixyj/epoll-shim#36 has been merged.
|
Thank you for this! Those changes look reasonable to me at a first glance. CheriBSD looks like a cool project! Is there a way for me to test it without a dev board? Is https://www.cl.cam.ac.uk/research/security/ctsrd/cheri/cheri-qemu.html the way to go? |
Yes, QEMU is the best way to get started. If you clone cheribuild, running |
|
And there is of course a target to build epoll-shim as well, |
|
Friendly ping :) |
|
ping? |
|
Sorry to ping again, but would it be possible to merge this? |
|
ping? |
arichardson
force-pushed
the
fixes-for-cheribsd
branch
from
090f859
to
f780fb7
Compare
|
rebased on latest master, @jiixyj any chance this could be merged? |
|
ping? |
|
ping? It would be nice if we could drop the local patch from CheriBSD ports |
|
@jiixyj Any chance this could be merged? |
|
Thanks for your patience -- I haven't forgotten about this! Some thoughts:
|
When running the tests on a CheriBSD system on Arm Morello, pointers are 16 bytes, so the implicit size of 8192 bytes is not large enough to hold all of struct atf_tc_impl_s_ and running the tests fails with a bounds violation. To fix this issue, this commit explicitly sizes the array instead of relying on the alignment and adds a static assertion to the source file to ensure this invariant holds. I've also reordered struct atf_tc_impl_s_ to avoid padding on CHERI systems.
…ents `fcntl()` abuses C variadic arguments to implement optional parameters. We want to forward this argument to the real `fcntl()` call, but doing so is non-portable. On most architectures, we can just read an `intptr_t`/`void*` and this will give us a (semi-)valid result even if no argument/an int was passed instead since `va_arg()` will generally read the next argument register or arbitrary data from the stack. On CHERI-enabled architectures, variadic arguments are tightly bounded, which means that reading a `void*` if an `int` was passed will result in a runtime trap. This commit adds a switch ensures that we read the correct number of bytes from the variadic arguments. This ugly workaround would not be needed if a vfcntl() call taking a va_list existed, or if fcntl just took an intptr_t argument. This reduces the number of test failures on CheriBSD Morello from 210 to 8.
Some systems such as 64-bit CHERI-RISC-V or Arm Morello have 128-bit pointers, so for them the sizeof(event) assertion needs to be updated.
This avoids a warning when building for CheriBSD (Morello purecap).
Trying to run the test executables during the build to get the list of test cases causes my cross-compilation to fail. Skip adding the test cases to CTest in that case (but still build them we can run them manually).
Due to a bug in CheriBSD, errno was set non-zero when starting this test, but without these checks the test confusingly failed in the later check which made me think something was wrong in timerfd_create. Check errno on startup to make this more obvious. See CTSRD-CHERI/cheribsd#1424
arichardson
force-pushed
the
fixes-for-cheribsd
branch
from
f780fb7
to
ecac389
Compare
|
Sorry for the delay - I've rebased the changes. Unfortunately there is currently no way to interpose fcntl more easily... |
I tried running the tests on my Morello board and noticed that most of them were failing. This patch series makes epoll-shim work correctly when running on CheriBSD (a fork of FreeBSD that adds support for CHERI-enabled architectures such as CHERI-RISC-V and Arm Morello).
With this patch series all but two tests pass for me and the last remaining failures will go away once CTSRD-CHERI/cheribsd#1424 has been fixed.