Support vhost-user-vsock with the QEMU VMM

[dimorinny]

The qemu_cli VMM only wired up the kernel's vhost-vsock device, so run_cvd's vsock servers fail in containers without AF_VSOCK. Emit a vhost-user-vsock-pci-non-transitional device backed by vhost_device_vsock instead. auto now keys off the guest target_arch and covers arm64 and riscv64, so x86_64 hosts running arm64 guests under qemu_cli switch to vhost-user vsock.

QEMU connects to the chardev once at startup instead of retrying like crosvm, so wait for vhost.socket to listen first. That wait reads /proc/net/unix instead of lsof, and must not connect: the backend would consume the probe as its one allowed frontend.

Bug: b/561671404
Test: cvd start --vm_manager=qemu_cli --vhost_user_vsock=true, for an
x86_64 guest on an x86_64 host and an arm64 guest on an x86_64
host. Both reach sys.boot_completed=1 with adb working over vsock.

[Databean]

and must not connect: the backend would consume the probe as its one allowed frontend.

That was the original motivation for using lsof, see aosp/3047934. Is there an issue using it now?

[Databean]

There are presubmit errors:

 FAIL: //cuttlefish/common/libs/utils:wait_for_unix_socket_clang_tidy (Exit 1) (see /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/execroot/_main/bazel-out/k8-fastbuild/testlogs/cuttlefish/common/libs/utils/wait_for_unix_socket_clang_tidy/test.log)
  ==================== Test output for //cuttlefish/common/libs/utils:wait_for_unix_socket_clang_tidy:
  37 warnings generated.
  INFO: From Testing //cuttlefish/common/libs/utils:wait_for_unix_socket_clang_tidy:
  warning: redefining builtin macro [clang-diagnostic-builtin-macro-redefined]
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:19:1: error: included header sched.h is not used directly [misc-include-cleaner,-warnings-as-errors]
     19 | #include <sched.h>
        | ^~~~~~~~~~~~~~~~~~
     20 | 
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:25:1: error: included header regex is not used directly [misc-include-cleaner,-warnings-as-errors]
     25 | #include <regex>
        | ^~~~~~~~~~~~~~~~
     26 | #include <sstream>
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:29:1: error: included header utility is not used directly [misc-include-cleaner,-warnings-as-errors]
     29 | #include <utility>
        | ^~~~~~~~~~~~~~~~~~
     30 | 
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:31:1: error: included header log.h is not used directly [misc-include-cleaner,-warnings-as-errors]
     31 | #include "absl/log/log.h"
        | ^~~~~~~~~~~~~~~~~~~~~~~~~
     32 | 
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:37:1: error: included header command.h is not used directly [misc-include-cleaner,-warnings-as-errors]
     37 | #include "cuttlefish/process/command.h"
        | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     38 | #include "cuttlefish/process/managed_stdio.h"
  /home/runner/.cache/bazel/_bazel_runner/a1ea073c0d21ddc5ae89652639f74223/sandbox/processwrapper-sandbox/10928/execroot/_main/cuttlefish/common/libs/utils/wait_for_unix_socket.cpp:38:1: error: included header managed_stdio.h is not used directly [misc-include-cleaner,-warnings-as-errors]
     38 | #include "cuttlefish/process/managed_stdio.h"
        | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
     39 | #include "cuttlefish/result/result.h"
  Suppressed 28 warnings (28 in non-user code).
  Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.
  6 warnings treated as errors
  ================================================================================

[dimorinny]

That was the original motivation for using lsof, see aosp/3047934. Is there an issue using it now?

lsof itself does the right check, but it's a problem in practice:

1. It isn't a declared dependency: cuttlefish-base doesn't depend on it, and container/image/Containerfile (debian:13, --no-install-recommends) doesn't pull it in, so the nightly cuttlefish-orchestration image has no /usr/bin/lsof. A missing lsof fails the boot (RunAndCaptureStdout rejects nonzero exits). Today that's hidden because the only callers are behind vhost_user_block (default off); this PR starts using the wait for vhost-user vsock, on by default for arm64/riscv64 guests under qemu_cli, so I expect lsof to start hitting problems.
2. lsof reports unix socket state only since 4.94.0 (Nov 2020). Ubuntu 22.04 and bullseye ship 4.93.2, where lsof -F TST prints no state line at all, so the wait times out after 30 s (verified on Ubuntu 22.04). That likely means vhost_user_block is already broken on those hosts.
3. Cost: ~0.7 s per call, since it scans every process (why [assemble_cvd] Replace slow lsof for file cleanup #2489 dropped it).

lsof 4.94+ reads the same thing anyway: /proc/net/unix Flags & __SO_ACCEPTCON. The column layout is documented in proc(5), and the kernel has printed that bit there unchanged since 2.6.12. The new code does that check in process, still without connecting, so the aosp/3047934 constraint holds.

Pushed:

• The vhost.socket wait is now QEMU-only, since crosvm already retries via --vhost-user-connect-timeout-ms.
• clang-tidy include fix: <sched.h> dropped, lsof-fallback-only headers under #ifndef linux.
• The flag check uses __SO_ACCEPTCON from <linux/net.h> instead of a literal.