Re: [PATCH] ipc: optimize semget/shmget/msgget for lots of keys

From: Guillaume Knispel
Date: Thu Aug 03 2017 - 13:14:46 EST


On Wed, Aug 02, 2017 at 01:06:44PM -0700, Davidlohr Bueso wrote:
> On Mon, 31 Jul 2017, Guillaume Knispel wrote:
> >static int __init ipc_init(void)
> >{
> >- sem_init();
> >- msg_init();
> >+ int err_sem, err_msg;
> >+
> >+ err_sem = sem_init();
> >+ WARN(err_sem, "ipc: sysV sem_init failed: %d\n", err_sem);
> >+ err_msg = msg_init();
> >+ WARN(err_msg, "ipc: sysV msg_init failed: %d\n", err_msg);
> > shm_init();
>
> This shows the ugliness of the underlying ipc init asymmetry. Specifically,
> 140d0b2108f (Do 'shm_init_ns()' in an early pure_initcall) was the final
> nail in the coffin to fix an exit_shm() race.
>
> While normally we could just initialize the ipc_ids fields statically and
> be over with initcall dependencies, your patch will require inits be done
> dynamically for the rhashtable_init(). Oh well.
>
> Also, why do you do this?
>
> >-pure_initcall(ipc_ns_init);
> >+core_initcall(ipc_ns_init);

In linux/init.h I saw that a pure_initcall is reserved to only
initialize variables and must have no dependency on anything else;
I interpreted that, + "pure" in the name, thinking we should not e.g.
allocate in a pure_initcall, however I see that net_ns_init() calls
kmem_cache_create() and others, so maybe we can keep ipc_ns_init() as
a pure_initcall?

Guillaume