ci: arm64 support #11
Conversation
|
@theihor do we still need this PR? |
@anakryiko if we are ok with using "unofficial" arm64 build of bpftrace, I can update this PR to add arm64 to CI. Let me know. |
please remind me what are the implications? will it break often? |
@anakryiko the arm64 bpftrace binary is not an officially supported release. So it is probably less tested. For this CI, getting it is not a simple wget from an authoritative source. Whenever we'll want to update it, it's a little bit more work than a version bump. One option (used in this PR for testing) is to create a fake release (like this one) whenever we want to update bpftrace. If it's not that important for bpftrace to be up to date, then I'd suggest to stick to the version that I tried. I don't expect much problems with it, and if I'm wrong about that we'll notice quickly. |
|
Yeah, freshness of bpftrace isn't that important, we don't use anything bleeding-edge or fancy here. Let's do the custom release and have arm64 support in CI, thanks! |
theihor
force-pushed
the
ci-arm64
branch
2 times, most recently
from
1aba64d to
e0bea81
Compare
theihor
force-pushed
the
ci-arm64
branch
5 times, most recently
from
b2601d3 to
d708a36
Compare
|
@anakryiko I had to fix a few things in awk magic, and disable struct-by-value test for arm too. See the commits for details. |
Rewrite install-bpftrace.sh script to download custom bpftrace-arm64 release [1] and use it for testing on aarch64. Add github-hosted arm64 runners to the test workflow matrix. [1] https://github.com/theihor/bpftrace/releases/tag/v0.22.1-arm64 Signed-off-by: Ihor Solodrai <[email protected]>
On arm64, argument spec of a tracepoint does not include dollar signs. In tests matching expected arguments exactly, add USDT specs for arm64. Signed-off-by: Ihor Solodrai <[email protected]>
Modify the condition for arg_types struct-by-value test case to only run on x86_64 with gcc. Signed-off-by: Ihor Solodrai <[email protected]>
Signed-off-by: Ihor Solodrai <[email protected]>
NF in awk relies on whitespace to split the "fields" in a line.
This lead to incorrectly counted argn in the following case:
Arguments: -4@[sp, 40]
Change how argn is computed: count /-?[0-9]+@/ matches in a
line (examples: '-4@[sp, 40]', '8@%rdi'),
Signed-off-by: Ihor Solodrai <[email protected]>
tests/fetch-usdts.awk
Outdated
| if (base != 0) { | ||
| base_key = filename ":" base; | ||
| if (!(base_key in basemap)) { | ||
| basemap[base_key] = ++base_cnt; |
There was a problem hiding this comment.
let's reduce nestedness:
base_key = filename ":" base;
if (base != 0 && !(base_key in basemap)) {
...
The base and semaphore offsets between different binaries may coincide, which leads to incorrect matching of USDTs by test scripts. Use filename + offset string to identify them for the purpose of USDT spec comparison. readelf -n prints the file path before dumping notes if more than one file is passed in, and this is now captured by fetch-usdts.awk. Signed-off-by: Ihor Solodrai <[email protected]>
No description provided.