[linux-vrf-core] [PATCH] locking too long -> panic

[Yon U. <hav...@gm...>]

Hi,

fib_add_vrf()
 {

   ... write_lock_bh(...);
     -> cpu_bh_disable(...)
     -> -> irq_stats.__local_bh_count++
  ...
     3x __fib_new_table();
     -> fib_hash_init
      -> -> kmalloc(, GFP_KERNEL) /*  allocs from the 256 bytes pool */
 }

eventually (aint randomness fun) the slab cache entry for 256 must grow,
and it hits:

  if (in_interrupt() && (flags & SLAB_LEVEL_MASK) != SLAB_ATOMIC)
                BUG();

in_interrupt() checks for "(... +irq_stats.__local_bh_count) != 0"

Anyway, now for my voodoo explanation of locking:

 On SMP the function is protected by lock_kernel()
(in inet_ioctl()), on UP no write access can happen, 'cos
this is user context and it is unpreemptable, and no BH will
write-access the values. So basically, only the last assigment
needs to be protected (not the "kmalloc"s ), and the check for
        if (fib_rules[vrf]) return -EEXIST;
is outside the protected region anyway.

enjoy,
 yon