[libc] fix scudo integration build

[SchrodingerZhu]

FYI, scudo is using our header while not using our implementation.

[llvmbot]

@llvm/pr-subscribers-libc

Author: Schrodinger ZHU Yifan (SchrodingerZhu)

Changes

FYI, scudo is using our header while not using our implementation.

---

Full diff: https://github.com/llvm/llvm-project/pull/116979.diff

1 Files Affected:

• (modified) libc/test/integration/scudo/CMakeLists.txt (+9)

diff --git a/libc/test/integration/scudo/CMakeLists.txt b/libc/test/integration/scudo/CMakeLists.txt
index a5f7e3b63d24c3..06d7ee6eca67d7 100644
--- a/libc/test/integration/scudo/CMakeLists.txt
+++ b/libc/test/integration/scudo/CMakeLists.txt
@@ -21,6 +21,15 @@ add_entrypoint_library(
     libc.src.errno.errno
     libc.src.unistd.__llvm_libc_syscall
     libc.src.sched.__sched_getcpucount
+    libc.src.unistd.sysconf
+    libc.src.sys.auxv.getauxval
+    libc.src.sys.prctl.prctl
+    libc.src.sys.mman.mmap
+    libc.src.sys.mman.munmap
+    libc.src.unistd.read
+    libc.src.unistd.write
+    libc.src.unistd.close
+    libc.src.fcntl.open
 )
 
 add_executable(

[michaelrj-google]

Thanks for fixing this! At some point I want to address the comment I left in this file and make this a proper integration test, so I'll file an issue for that.

[nickdesaulniers]

Thanks for figuring this out! It had me stumped yesterday. Can you tell me a little more about how you ended up figuring this out?

[SchrodingerZhu]

@nickdesaulniers I am not sure why it does not reproduce on your side, but it just reproduces and the debugging story starts there.

So I just attached the debugger and found out that it errors out on a failure of mmap. Using strace, I can see that the failed mmap is to apply a MAP_FIXED on an address that is not page-aligned.

Then, it just comes to me that the allocator may get the wrong page size.

I re-run the test with debugger attached, and I tried sysconf(1) in the debugger (_SC_PAGESIZE=1 in our header), and it returned a value that is apparently not correct.

So, it must be the case that the implementation does not match the header. Indeed, the test can be fixed by enforcing the symbol implementation from our side.

---

My guess: the reason why it does not reproduce may be related to how header to inclusion is passed to scudo. Maybe @lntue and you are not using fullbuild?

[nickdesaulniers]

Ah! I was reproducing on the buildbot, and was attaching gdb to spot a malloc that would fail.

Breakpoint 1, __GI___mmap64 (addr=0xffd7f7ee77ca, len=262144, prot=3, flags=50, fd=-1, offset=0)

I guess I didn't spot that 0xffd7f7ee77ca is not a multiple of the page size and that 50 is MAP_FIXED.

Anyways, bravo, nice work!