| Message ID | 20251022194628.2462629-1-skandigraun@gmail.com |
|---|---|
| State | Accepted, archived |
| Commit | 81f7b60fb1c5096bbc233f632040d1ea9ec5bb21 |
| Headers | show |
| Series | tar: use diffutils for ptest instead of busybox | expand |
On 22 Oct 2025, at 20:46, Gyorgy Sarvari via lists.openembedded.org <skandigraun=gmail.com@lists.openembedded.org> wrote:
> +RDEPENDS:${PN}-ptest += “diffutils"
From the experience of looking at assignments like this in a few years time and trying to explain why they exist; can you add a one-line summary of your commit message as a comment here?
(also, curious how busybox fmt is so slow compared to diffutils, considering “read files until they differ” isn’t exactly rocket science…)
Thanks,
Ross
On 10/27/25 18:38, Ross Burton wrote: > On 22 Oct 2025, at 20:46, Gyorgy Sarvari via lists.openembedded.org <skandigraun=gmail.com@lists.openembedded.org> wrote: >> +RDEPENDS:${PN}-ptest += “diffutils" > From the experience of looking at assignments like this in a few years time and trying to explain why they exist; can you add a one-line summary of your commit message as a comment here? Will do. > (also, curious how busybox fmt is so slow compared to diffutils, considering “read files until they differ” isn’t exactly rocket science…) Busybox[1] reads the files in sizeof(int) chunks, and compares them as ints. diffutils[2] reads "bufsize" amount each time, and does memcmp on them. Haven't looked into the exact size of bufsize, but at the first sight it looks like it's the blocksize of the device being read, which is most likely bigger than 4. Haven't done a benchmark on these lines, but Busybox's version looks more KISS-y rather than speedy. [1]: https://elixir.bootlin.com/busybox/1.37.0/source/editors/cmp.c#L122 [2]: https://cgit.git.savannah.gnu.org/cgit/diffutils.git/tree/src/cmp.c#n487 > > Thanks, > Ross
diff --git a/meta/recipes-extended/tar/tar_1.35.bb b/meta/recipes-extended/tar/tar_1.35.bb index ea0993a909..206e084557 100644 --- a/meta/recipes-extended/tar/tar_1.35.bb +++ b/meta/recipes-extended/tar/tar_1.35.bb @@ -94,3 +94,5 @@ BBCLASSEXTEND = "native nativesdk" # Avoid false positives from CVEs in node-tar package # For example CVE-2021-{32803,32804,37701,37712,37713} CVE_PRODUCT = "gnu:tar" + +RDEPENDS:${PN}-ptest += "diffutils"
A testcase (sparse03) sometimes times out on the AB, in qemu (without kvm): the test generates an 8GB sparse file, tars it, untars it, and then it compares the two versions with cmp. This process, going through 16GB of data (using one thread, with cmp) takes some time anyway, but when there is extra load on the host machine, and qemu can't use its core exclusively, then it can take more than 5 minutes easily (which is the default ptest timeout). However the full version of cmp from diffutils seems to be more efficient than the busybox version: When using busybox on my idle machine (w/ qemuriscv64) the test case execution takes 150s, and it almost always times out when there is extra load. Using diffutils, my idle machine executes the same testcase in 55s, and it never times out even if there is high load on the host system (execution always stayed under 3 minutes). Due to this switch to diffutils when running ptest. Fixes [YOCTO 15884] Signed-off-by: Gyorgy Sarvari <skandigraun@gmail.com> --- meta/recipes-extended/tar/tar_1.35.bb | 2 ++ 1 file changed, 2 insertions(+)