arch/sim: save errno in sim context switch to avoid coroutine make errno mess between host api calls.

[jasonbu]

Summary

As discussion in #16742, manually up_irq_save is not a good idea, we should save errno in "interrupt" to ensure the errno use as usual will more easy to mantain.

This patch add errno in sim context, and save it before call coroutine setjump, restore before longjump.

When there is lot of sim feature provide by host, the errno may change by coroutine/switch context/...
The code to reproduce issue:
paste it into apps/examples/hello/hello_main.c

#include <nuttx/config.h>
#include <stdio.h>
#include <unistd.h>
#include <spawn.h>
#include <fcntl.h>
#include <sys/wait.h>
#include <sys/stat.h>

void *thread_gen_errno(void *arg)
{
  bool *gen_done = (bool *)arg;
  const char *dir_path = "/data1/test_is_dir";
  const char *chdir_path = "/data1";
  int ret = 0;

  ret = mkdir(dir_path, 0777);
  printf("mkdir: %s done, %d\n", chdir_path, ret);

  printf("child errno thread start\n");
  while(true) {
      usleep(1);

      if (*gen_done) {
        /* Regular exit */
        ret = 0;
        break;
      }

extern void host_errno_set(int errcode);
extern int host_errno_get(void);
      host_errno_set(EISDIR);
      ret = host_errno_get();
      if (ret != EISDIR) {
        printf("host_errno_set/get fail: set EISDIR, get %d\n", ret);
        *gen_done = true;
        break;
      }
    }

  rmdir(dir_path);

  printf("child errno thread end\n");
  return (void *)(uintptr_t)ret;
}

int main(int argc, FAR char *argv[])
{
  const int maxtimes = 1000*10000;
  const char *dir_path = "/data/test_rm_dir";
  int times = 0;
  bool gen_done = false;
  int ret;
  pid_t tid;
  pthread_attr_t attr;

  pthread_attr_init(&attr);
  attr.priority = CONFIG_EXAMPLES_HELLO_PRIORITY + 1;

  ret = pthread_create(&tid, &attr, thread_gen_errno, &gen_done);
  pthread_attr_destroy(&attr);

  if (ret != 0) {
      printf("Failed to spawn child thread: %d\n", ret);
      return -1;
    }

  for (int i = 0; i < maxtimes; i++) {
      times ++;
      ret = rmdir(dir_path);

      if (ret == 0) {
          printf("rmdir: %s succ", dir_path);
      } else {
        int err = errno;

        if (err == ENOENT) {
          /* printf("dir not exist\n"); The expected err */
        } else {
          printf("rmdir fail: %s (errno: %d)\n", strerror(err), err);
          break;
        }
      }

      if (gen_done) {
        /* Child thread exit abnormally */
        break;
      }

      if (times % 100000 == 0) {
        printf("exec done %d times\n", times);
      }
    }

  gen_done = true;
  pthread_join(tid, (void *)&ret);

  if (times == maxtimes && ret == 0) {
    printf("rmdir test success, total exec %d times\n", times);
  } else {
    printf("break after %d times, ret: %d\n", times, (int)ret);
  }

  return 0;
}

Impact

Now we don't have to take care the errno changed when "interrupt" happened. the context switch will matain the errno correctly.
Code to reproduce will no longer report break.

Testing

CI test, ubuntu 24.04 sim/nsh, local complex sim project.
After revert ac5b38c and add a host_errno_get & host_errno_set.
These demo will able to re-produce it in sim/nsh.

rmdir fail: Unknown error 21 (errno: 21)
child errno thread end
break after 3579966 times, ret: 0

and succ after whole pr.

rmdir test success, total exec 10000000 times

[xiaoxiang781216]

@yamt could you review whether this patchset fix your concern?

[cederom]

Thank you @jasonbu :-) Do we have some sort of nuke test application for this case where corner cases can be generated and verified? It may be good to see how things are handled on various Sim:OS (i.e. Linux, BSD, macOS, etc)?

Is the code in PR description enough / sufficient to cover all scenarios or just a starting point? Can we put that as starting point of nuttx-apps/testing/simtest and create sim:simtest target so its easy to build/verify for everyone? :-)

[jasonbu]

Thank you @jasonbu :-) Do we have some sort of nuke test application for this case where corner cases can be generated and verified? It may be good to see how things are handled on various Sim:OS (i.e. Linux, BSD, macOS, etc)?

Is the code in PR description enough / sufficient to cover all scenarios or just a starting point? Can we put that as starting point of nuttx-apps/testing/simtest and create sim:simtest target so its easy to build/verify for everyone? :-)

There is some dependency if reproduce with this code,

• sim and /data as hostfs
• have other host business working with polling designed feature.

so in my environment, sim/citest cannot reproduce, but in internal project can reproduce with less than 10000 times loop.

if you have time to ensure it reproduced and add to sim-testing will appreciate

[jasonbu]

Thank you @jasonbu :-) Do we have some sort of nuke test application for this case where corner cases can be generated and verified? It may be good to see how things are handled on various Sim:OS (i.e. Linux, BSD, macOS, etc)?

Is the code in PR description enough / sufficient to cover all scenarios or just a starting point? Can we put that as starting point of nuttx-apps/testing/simtest and create sim:simtest target so its easy to build/verify for everyone? :-)

hi, @cederom ,
rebase to the lastest code, still working, and updated the sample.
it still corner case, and not eary to take it into standard test/example.

The 1 million times test, will not able to reproduce the problem.
using 10 million times, will >50% re-produce. also try use 2nd mount point and 2nd host relative API to generate a host-API caused errno set.
The possiblily will even lower, so finally use host_errno_set and host_errno_get to make it more eary to re-produce the problem.

In the real-complex scene. True issue will be, if we don't use host_uninterruptible_errno, will cause errno randomly covered by other host-API. And if we depends errno to do host behavior(should be regular desisng). errno occupation will cause a mess.

CC @yamt

[cederom]

Thank you @jasonbu :-)

• Linker issue https://github.com/apache/nuttx/actions/runs/20910485474/job/60072337029?pr=16795#step:11:480
• Mixed case were here and are not introduced by the change.

[cederom]

Thank you @jasonbu for the fix! Let's just make sure CI build errors are fixed before merge :-)

[cederom]

TODO: Linker error fix :-)

[jasonbu]

TODO: Linker error fix :-)

Ths for review, and help find where CI report error, the host_uninterruptible_errno usage increased during these days. fixed, and please check.

fixed