Re: [PATCH v4 2/2] mm: Init page count in reserve_bootmem_region when MEMINIT_EARLY

[Yajun Deng]

On 2023/10/13 16:48, Mike Rapoport wrote:
> On Thu, Oct 12, 2023 at 05:53:22PM +0800, Yajun Deng wrote:
> > On 2023/10/12 17:23, David Hildenbrand wrote:
> > > On 10.10.23 04:31, Yajun Deng wrote:
> > > > On 2023/10/8 16:57, Yajun Deng wrote:
> > > > > > That looks wrong. if the page count would by pure luck be 0
> > > > > > already for hotplugged memory, you wouldn't clear the reserved
> > > > > > flag.
> > > > > >
> > > > > > These changes make me a bit nervous.
> > > > > Is 'if (page_count(page) || PageReserved(page))' be safer? Or do I
> > > > > need to do something else?
> > > > How about the following if statement? But it needs to add more patch
> > > > like v1 ([PATCH 2/4] mm: Introduce MEMINIT_LATE context).
> > > >
> > > > It'll be safer, but more complex. Please comment...
> > > >
> > > >      if (context != MEMINIT_EARLY || (page_count(page) ||
> > > > PageReserved(page)) {
> > > Ideally we could make initialization only depend on the context, and not
> > > check for count or the reserved flag.
> > This link is v1,
> > https://lore.kernel.org/all/20230922070923.355656-1-yajun.deng@xxxxxxxxx/
> >
> > If we could make initialization only depend on the context, I'll modify it
> > based on v1.
> Although ~20% improvement looks impressive, this is only optimization of a
> fraction of the boot time, and realistically, how much 56 msec saves from
> the total boot time when you boot a machine with 190G of RAM?


There are a lot of factors that can affect the total boot time. 56 msec 
saves may be insignificant.

But if we look at the boot log, we'll see there's a significant time jump.

before:

[    0.250334] ACPI: PM-Timer IO Port: 0x508

[    0.618994] Memory: 173413056K/199884452K available (18440K kernel 
code, 4204K rwdata, 5608K rodata, 3188K init, 17024K bss, 5499616K 
reserved, 20971520K cma-reserved)


after:

[    0.260229] software IO TLB: area num 32.

[    0.563497] Memory: 173413056K/199884452K available (18440K kernel 
code, 4204K rwdata, 5608K rodata, 3188K init, 17024K bss, 5499616K 
reserved, 20971520K cma-reserved)


Memory initialization is time consuming in the boot log.

> I still think the improvement does not justify the churn, added complexity
> and special casing of different code paths of initialization of struct pages.
>   


Because there is a loop, if the order is MAX_ORDER, the loop will run 
1024 times. The following 'if' would be safer:

'if (context != MEMINIT_EARLY || (page_count(page) || >> 
PageReserved(page)) {'

This is a foundation. We may change this when we are able to safely 
remove page init in the hotplug context one day.

So the case for the early context is only temporary.

> > @Mike,  By the way,  this code will cost more time:
> >
> >                  if (context == MEMINIT_HOTPLUG)
> >                          flags = INIT_PAGE_COUNT | INIT_PAGE_RESERVED;
> >                  __init_single_page(page, pfn, zone, nid, flags);
> >
> >
> > [    0.014999] On node 0, zone DMA32: 31679 pages in unavailable ranges
> > [    0.311560] ACPI: PM-Timer IO Port: 0x508
> >
> >
> > This code will cost less time:
> >
> >                  __init_single_page(page, pfn, zone, nid, 0);
> >                  if (context == MEMINIT_HOTPLUG) {
> >                          init_page_count(page);
> >                          __SetPageReserved(page);
> >
> > [    0.014299] On node 0, zone DMA32: 31679 pages in unavailable ranges
> > [    0.250223] ACPI: PM-Timer IO Port: 0x508