Re: [PATCH] mm/mm_init.c: remove spinlock in early_pfn_to_nid()

[Mike Rapoport]

On Thu, Jun 15, 2023 at 03:02:58AM +0000, Yajun Deng wrote:
> June 14, 2023 7:53 PM, "Mike Rapoport" <rppt@xxxxxxxxxx> wrote:
> 
> > Hi,
> > 
> > On Wed, Jun 14, 2023 at 11:28:32AM +0000, Yajun Deng wrote:
> > 
> >> June 14, 2023 7:09 PM, "Greg KH" <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> >> 
> >> On Wed, Jun 14, 2023 at 07:03:24PM +0800, Yajun Deng wrote:
> >> 
> >> When the system boots, only one cpu is enabled before smp_init().
> >> So the spinlock is not needed in most cases, remove it.
> >> 
> >> Add spinlock in get_nid_for_pfn() because it is after smp_init().
> >> 
> >> So this is two different things at once in the same patch?
> >> 
> >> Or are they the same problem and both need to go in to solve it?
> >> 
> >> And if a spinlock is not needed at early boot, is it really causing any
> >> problems?
> >> 
> >> They are the same problem.
> >> I added pr_info in early_pfn_to_nid(), found get_nid_for_pfn() is the only
> >> case need to add spinlock.
> >> This patch tested on my x86 system.
> > 
> > Are you sure it'll work on !x86?
> >
> 
> I'm probably sure of that, although I don't have a !x86 machine.
> 
> early_pfn_to_nid() is called in smp_init() and kasan_init() on 
> different architectures. If it works well on x86, it'll work on
> !x86.

This is often not true. Please verify that other architectures do not call
early_pfn_to_nid() after smp_init(). The explanation why it is safe should
be a part of the changelog.

> >> Signed-off-by: Yajun Deng <yajun.deng@xxxxxxxxx>
> >> ---
> >> drivers/base/node.c | 11 +++++++++--
> >> mm/mm_init.c | 18 +++---------------
> >> 2 files changed, 12 insertions(+), 17 deletions(-)
> >> 
> >> diff --git a/drivers/base/node.c b/drivers/base/node.c
> >> index 9de524e56307..844102570ff2 100644
> >> --- a/drivers/base/node.c
> >> +++ b/drivers/base/node.c
> >> @@ -748,8 +748,15 @@ int unregister_cpu_under_node(unsigned int cpu, unsigned int nid)
> >> static int __ref get_nid_for_pfn(unsigned long pfn)
> >> {
> >> #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
> >> - if (system_state < SYSTEM_RUNNING)
> >> - return early_pfn_to_nid(pfn);
> >> + static DEFINE_SPINLOCK(early_pfn_lock);
> >> + int nid;
> >> +
> >> + if (system_state < SYSTEM_RUNNING) {
> >> + spin_lock(&early_pfn_lock);
> >> + nid = early_pfn_to_nid(pfn);
> >> + spin_unlock(&early_pfn_lock);
> >> 
> >> Adding an external lock for when you call a function is VERY dangerous
> >> as you did not document this anywhere, and there's no way to enforce it
> >> properly at all.
> >> 
> >> I should add a comment before early_pfn_to_nid().
> >> 
> >> Does your change actually result in any boot time changes? How was this
> >> tested?
> >> 
> >> Just a bit.
> > 
> > Just a bit tested? Or just a bit of boot time changes?
> > For the latter, do you have numbers?
> > 
> 
> For the latter, the most beneficial function is memmap_init_reserved_pages(),
> the boot time changes depending on whether DEFERRED_STRUCT_PAGE_INIT
> is defined or not. 
> 
> -->memmap_init_reserved_pages()
>    -->for_each_reserved_mem_range()
>       reserve_bootmem_region()
>       -->for()
>          init_reserved_page()
>          --> early_pfn_to_nid()
 
A better solution would be to pass nid to reserve_bootmem_range() and drop
the call to early_pfn_to_nid() in init_reserved_page(). 

Then there won't be lock contention and no need for fragile changes in the
locking.

> If define CONFIG_DEFERRED_STRUCT_PAGE_INIT:
> 
> before:
> memmap_init_reserved_pages()   1.87 seconds
> after:
> memmap_init_reserved_pages()   1.27 seconds
> 
> 32% time reduction.

These measurements should be part of the changelog. 
 
> If not define CONFIG_DEFERRED_STRUCT_PAGE_INIT:
> 
> early_pfn_to_nid() is called by few, 
> boot time didn't change.
> 
> 
> By the way, this machine has 190GB RAM.
> 
> > --
> > Sincerely yours,
> > Mike.

-- 
Sincerely yours,
Mike.