add gemma accelerator classification results

1 files changed, 8 insertions, 0 deletions

diff --git a/results/classifier/accel-gemma3:12b/tcg/1875702 b/results/classifier/accel-gemma3:12b/tcg/1875702
new file mode 100644
index 000000000..b29ee7b07
--- /dev/null
+++ b/results/classifier/accel-gemma3:12b/tcg/1875702
@@ -0,0 +1,8 @@
+
+madvise reports success, but doesn't implement WIPEONFORK.
+
+The implementation of madvise (linux-user/syscall.c:11331, tag v5.0.0-rc4) always returns zero (i.e. success). However, an application requesting (at least) MADV_WIPEONFORK may need to know whether the call was actually successful. If not (because the kernel doesn't support WIPEONFORK) then it will need to take other measures to provide fork-safety (such as drawing entropy from the kernel in every case). But, if the application believes that WIPEONFORK is supported (because madvise returned zero), but it actually isn't (as in qemu), then it may forego those protections on the assumption that WIPEONFORK will provide fork-safety.
+
+Roughly, the comment in qemu that says "This is a hint, so ignoring and returning success is ok." is no longer accurate in the presence of MADV_WIPEONFORK.
+
+(This is not purely academic: BoringSSL is planning on acting in this way. We found the qemu behaviour in pre-release testing and are planning on making an madvise call with advice=-1 first to test whether unknown advice values actually produce EINVAL.)
\ No newline at end of file