Re: 174cc7187e6f ACPICA: Tables: Back port acpi_get_table_with_size() and early_acpi_os_unmap_memory() from Linux kernel

[Paul E. McKenney]

On Tue, Jan 10, 2017 at 02:27:16AM +0100, Rafael J. Wysocki wrote:
> On Tue, Jan 10, 2017 at 12:52 AM, Borislav Petkov <bp@xxxxxxxxx> wrote:
> > On Tue, Jan 10, 2017 at 12:40:39AM +0100, Borislav Petkov wrote:
> >> Lemme run it.
> >
> > Well, it boots but I get:
> >
> > [    0.291447] ------------[ cut here ]------------
> > [    0.291702] WARNING: CPU: 0 PID: 1 at kernel/rcu/tree.c:3993 rcu_scheduler_starting+0x5c/0x70
> > [    0.292107] Modules linked in:
> > [    0.292277] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.10.0-rc3+ #21
> > [    0.292540] Hardware name: HP HP EliteBook 745 G3/807E, BIOS N73 Ver. 01.08 01/28/2016
> > [    0.292893] Call Trace:
> > [    0.293072]  ? dump_stack+0x46/0x63
> > [    0.293285]  ? __warn+0xec/0x110
> > [    0.293487]  ? rcu_scheduler_starting+0x5c/0x70
> > [    0.293735]  ? kernel_init_freeable+0x58/0x19a
> > [    0.293976]  ? rest_init+0x80/0x80
> > [    0.294153]  ? kernel_init+0xa/0x100
> > [    0.294334]  ? ret_from_fork+0x22/0x30
> > [    0.294525] ---[ end trace 4c0fe009ed4dc740 ]---
> >
> > TBH, I like Rafael's suggestion in the other mail to stick with fixing
> > this in ACPI, especially this is an ACPI problem, not RCU. Well,
> > more or less: RCU could be taught to *not* do schedule_work() if
> > workqueue_init() hasn't happened yet but that's a tangential.
> >
> > So, I'm going to bed. When I wake up, I want to see working fixes!
> >
> > :-)))
> 
> Well, if the https://patchwork.kernel.org/patch/9504277/ patch from Lv
> worked, the attached one should work too (please test), but it can be
> justified in a slightly more convincing way.
> 
> Namely, the idea is that acpi_os_read/write_memory() should never be
> used before invoking acpi_os_initialize() and since those are the only
> places where the list of memory regions is walked under RCU without
> extra locking, it is safe to skip the RCU synchronization until that
> happens.
> 
> Thanks,
> Rafael

Makes sense to me!

It looks like I can make the grace-period-free boot-time window
for CONFIG_PREEMPT=y kernels quite a bit narrower, but it does not
look like something suitable for jamming into 4.10.

							Thanx, Paul

> ---
>  drivers/acpi/osl.c |    8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> Index: linux-pm/drivers/acpi/osl.c
> ===================================================================
> --- linux-pm.orig/drivers/acpi/osl.c
> +++ linux-pm/drivers/acpi/osl.c
> @@ -378,7 +378,9 @@ static void acpi_os_drop_map_ref(struct
>  static void acpi_os_map_cleanup(struct acpi_ioremap *map)
>  {
>  	if (!map->refcount) {
> -		synchronize_rcu_expedited();
> +		if (acpi_os_initialized)
> +			synchronize_rcu_expedited();
> +
>  		acpi_unmap(map->phys, map->virt);
>  		kfree(map);
>  	}
> @@ -671,6 +673,8 @@ acpi_os_read_memory(acpi_physical_addres
>  	bool unmap = false;
>  	u64 dummy;
>  
> +	WARN_ON_ONCE(!acpi_os_initialized);
> +
>  	rcu_read_lock();
>  	virt_addr = acpi_map_vaddr_lookup(phys_addr, size);
>  	if (!virt_addr) {
> @@ -716,6 +720,8 @@ acpi_os_write_memory(acpi_physical_addre
>  	unsigned int size = width / 8;
>  	bool unmap = false;
>  
> +	WARN_ON_ONCE(!acpi_os_initialized);
> +
>  	rcu_read_lock();
>  	virt_addr = acpi_map_vaddr_lookup(phys_addr, size);
>  	if (!virt_addr) {