On 11/09/2017 07:43 PM, Thomas Gleixner wrote:
> On Sun, 5 Nov 2017, Prarit Bhargava wrote:
>> [v5]: Change kmalloc to GFP_ATOMIC to fix "sleeping function" warning on
>> virtual machines.
>
> What has this to do with virtual machines? The very same issue is on
> physcial hardware because this is called from the early CPU bringup code
> with interrupts and preemption disabled.
There was a Intel test bot report of a failure during boot on virtual systems
with Andi's patch.
>
>> + /* Allocate and copy a new array */
>> + ltp_pkg_map_new = kmalloc(logical_packages * sizeof(u16), GFP_ATOMIC);
>> + BUG_ON(!ltp_pkg_map_new);
>
> Having an allocation in that code path is a bad idea. First of all the
> error handling in there is just crap, because the only thing you can do is
> panic. Aside of that atomic allocations should be avoided when we can and
> we can.
>
> Sorry I missed that when looking at the patch earlier. Something along this
> makes it work proper:
>
> struct pkg_map {
> unsigned int size;
> unsigned int used;
> unsigned int map[0];
> };
>
> static struct pkg_map *logical_to_physical_pkg_map __read_mostly;
>
> static int resize_pkg_map(void)
> {
> struct pkg_map *newmap, *oldmap = logical_to_physical_pkg_map;
> int size;
>
> if (oldmap->size > oldmap->used)
> return 0;
>
> size = sizeof(*oldmap) + sizeof(unsigned int) * oldmap->size;
> newmap = kzalloc(size + sizeof(unsigned int));
> if (!newmap)
> return -ENOMEM;
>
> memcpy(newmap, oldmap, size);
> newmap->size++;
> logical_to_physical_pkg_map = newmap;
> kfree(oldmap);
> return 0;
> }
>
> int __cpu_up(....)
> {
> if (resize_pkg_map())
> return -ENOMEM;
> return smp_ops.cpu_up(....);
> }
>
> static void update_map(....)
> {
> if (find_map())
> return;
> map->map[map->used] = physid;
> map->used++;
> }
>
> static void smp_init_package_map()
> {
> struct pkg_map *map;
>
> map = kzalloc(sizeof(*newmap) + sizeof(unsigned int));
> map->size = 1;
> }
>
> See? No BUG_ON() in the early secondary cpu boot code. If memory allocation
> fails the thing goes back gracefully.
>
> Locking/barriers omitted as you have choices here:
>
> 1) RCU
>
> Needs the proper RCU magic for the lookup and the pointer swap.
>
> That requires also a proper barrier between the assignement of the
> new id and the increment of the used count plus the corresponding one
> on the read side.
>
> 2) mutex
>
> Must be held when swapping the pointers and across lookup
>
> Same barrier requirement as RCU
>
> 3) raw_spinlock
>
> Must be held when swapping the pointers and across lookup
>
> No barriers as long as you hold the lock across the assignement and
> increment.
>
> All of that works. There is no way to make sure that a lookup is fully
> serialized against a concurrent update. Even if the lookup holds
> cpu_read_lock() the new package might arrive right after the unlock.
>
Thanks Thomas.
Andi, do you want to take a look at this?
P.
> Thanks,
>
> tglx
>
