Attempt to detect QEMU hangs

When building cross platform images with Docker,
QEMU is often used under the hood and can have a bug
that cause forked processes to deadlock.

Before spawning workers we test for that bug.

Fix: #495
Closes: #497

Co-Authored-By: Sarun Rattanasiri <[email protected]>

Author: byroot

CHANGELOG.md

@@ -1,5 +1,7 @@
 # Unreleased
 
+* Attempt to detect a QEMU bug that can cause `bootsnap precompile` to hang forever when building ARM64 docker images
+  from x86_64 machines. See #495.
 * Improve CLI to detect cgroup CPU limits and avoid spawning too many worker processes.
 
 # 1.18.4

lib/bootsnap/cli/worker_pool.rb

@@ -2,6 +2,7 @@
 
 require "etc"
 require "rbconfig"
+require "io/wait" unless IO.method_defined?(:wait_readable)
 
 module Bootsnap
   class CLI
@@ -17,12 +18,19 @@ def create(size:, jobs:)
         end
 
         def default_size
-          size = [Etc.nprocessors, cpu_quota || 0].min
+          nprocessors = Etc.nprocessors
+          size = [nprocessors, cpu_quota || nprocessors].min
           case size
           when 0, 1
             0
           else
-            size
+            if fork_defunct?
+              $stderr.puts "warning: faulty fork(2) detected, probably in cross platform docker builds. " \
+                           "Disabling parallel compilation."
+              0
+            else
+              size
+            end
           end
         end
 
@@ -32,6 +40,7 @@ def cpu_quota
               # cgroups v2: https://docs.kernel.org/admin-guide/cgroup-v2.html#cpu-interface-files
               cpu_max = File.read("/sys/fs/cgroup/cpu.max")
               return nil if cpu_max.start_with?("max ") # no limit
+
               max, period = cpu_max.split.map(&:to_f)
               max / period
             elsif File.exist?("/sys/fs/cgroup/cpu,cpuacct/cpu.cfs_quota_us")
@@ -40,11 +49,40 @@ def cpu_quota
               # If the cpu.cfs_quota_us is -1, cgroup does not adhere to any CPU time restrictions
               # https://docs.kernel.org/scheduler/sched-bwc.html#management
               return nil if max <= 0
+
               period = File.read("/sys/fs/cgroup/cpu,cpuacct/cpu.cfs_period_us").to_f
               max / period
             end
           end
         end
+
+        def fork_defunct?
+          return true unless ::Process.respond_to?(:fork)
+
+          # Ref: https://github.com/Shopify/bootsnap/issues/495
+          # The second forked process will hang on some QEMU environments
+          r, w = IO.pipe
+          pids = 2.times.map do
+            ::Process.fork do
+              exit!(true)
+            end
+          end
+          w.close
+          r.wait_readable(1) # Wait at most 1s
+
+          defunct = false
+
+          pids.each do |pid|
+            _pid, status = ::Process.wait2(pid, ::Process::WNOHANG)
+            if status.nil? # Didn't exit in 1s
+              defunct = true
+              Process.kill(:KILL, pid)
+              ::Process.wait2(pid)
+            end
+          end
+
+          defunct
+        end
       end
 
       class Inline