[MESOS-4577] libprocess can not run on 16-byte aligned stack mandatory architecture(aarch64)

• Type: Bug
• Status: Accepted
• Priority: Major
• Resolution: Unresolved
• Component/s: libprocess, stout
• Labels: mesosphere

Environment

Linux 10-175-112-202 4.1.6-rc3.aarch64 #1 SMP Mon Oct 12 01:43:03 UTC 2015 aarch64 aarch64 aarch64 GNU/Linux

• Flags: Patch, Important
• Assignee: AndyPang
• Reporter: AndyPang
• Shepherd: Jie Yu
• Created: 02/Feb/16 02:48
• Updated: 26/Nov/18 13:36

Description

mesos run in AArch64 will get error, the log is:

E0101 00:06:56.636520 32411 slave.cpp:3342] Container 'b6be429a-08f0-4d52-b01d-abfcb6e0106b' for executor 'hello.84d205ae-f626-11de-bd66-7a3f6cf980b9' of framework '868b9f04-9179-427b-b050-ee8f89ffa3bd-0000' failed to start: Failed to fork executor: Failed to clone child process: Failed to clone: Invalid argument 

the "clone" achieve in libprocess 3rdparty stout library(in linux.hpp) packaging a syscall "clone" :

clone

inline pid_t clone(const lambda::function<int()>& func, int flags)
{
  // Stack for the child.
  // - unsigned long long used for best alignment.
  // - 8 MiB appears to be the default for "ulimit -s" on OSX and Linux.
  //
  // NOTE: We need to allocate the stack dynamically. This is because
  // glibc's 'clone' will modify the stack passed to it, therefore the
  // stack must NOT be shared as multiple 'clone's can be invoked
  // simultaneously.
  int stackSize = 8 * 1024 * 1024;
  unsigned long long *stack =
    new unsigned long long[stackSize/sizeof(unsigned long long)];

  pid_t pid = ::clone(
      childMain,
      &stack[stackSize/sizeof(stack[0]) - 1],  // stack grows down.
      flags,
      (void*) &func);

  // If CLONE_VM is not set, ::clone would create a process which runs in a
  // separate copy of the memory space of the calling process. So we destroy the
  // stack here to avoid memory leak. If CLONE_VM is set, ::clone would create a
  // thread which runs in the same memory space with the calling process.
  if (!(flags & CLONE_VM)) {
    delete[] stack;
  }

  return pid;
}

syscal "clone" parameter stack is 8-byte aligned,so if in 16-byte aligned stack mandatory architecture(aarch64) it will get error.

Issue Links

• duplicates: MESOS-6835 Fix SIGBUS crash on ARM64/AArch64