Suka,
Thanks for the report.
Sukadev Bhattiprolu wrote:
> Probably premature :-) but tried to C/R a simple threaded application
> (running as container-init).
>
> First got an -EINVAL due to following check in may_checkpoint_task():
>
> /*
> * FIX: for now, disallow siblings of container init created
> * via CLONE_PARENT (unclear if they will remain possible)
> */
> if (ctx->root_init && t != ctx->root_task &&
> t->real_parent == ctx->root_task->real_parent)
>
> Assuming we are unintentionally excluding CLONE_THREAD with the
> above check, I added a check for tgid:
>
> if (ctx->root_init && t != ctx->root_task &&
> t->real_parent == ctx->root_task->real_parent &&
> t->tgid != ctx->root_task->tgid) {
>
> This got past the -EINVAL but the test failed the ckpt_obj_contained() check.
Yes, I see no reason to prevent multi-threaded container init.
Will add this to the ckpt-v15-dev git tree.
>
> c/r: FILE users 2 != count 6 objref 9
>
> The main-thread opened a single file (log file). The other threads don't
> write to it (yet). The count '6' corresponds to the number of threads in
> the application.
>
> I suspect that C/R code is incrementing obj->users once per thread for
> the log file even though the threads share the file_struct reference.
> (pthread_create() sets CLONE_FILES so the file_struct is shared between
> threads).
>
Indeed, the current code doesn't yet handle files_struct as a
shared object, so anything with CLONE_FILES isn't done correctly.
That's on the todo-list...
Oren.
> Will post my test programs to Serge's new git-tree next week.
>
> Sukadev
>
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/containers
