Re: [PATCH] x86, mm: Reserver some memory for bootmem allocator for NO_BOOTMEM

[Feng Tang]

Hi Thomas,

On Thu, Aug 30, 2018 at 03:05:31PM +0200, Thomas Gleixner wrote:
> On Thu, 30 Aug 2018, Feng Tang wrote:
> > On Thu, Aug 30, 2018 at 02:49:15PM +0200, Michal Hocko wrote:
> > > On Thu 30-08-18 13:54:19, Thomas Gleixner wrote:
> > > > On Thu, 30 Aug 2018, Michal Hocko wrote:
> > > > > On Thu 30-08-18 12:44:02, Peter Zijlstra wrote:
> > > > > > On Thu, Aug 30, 2018 at 05:03:19PM +0800, Feng Tang wrote:
> > > > > > > The root cause is that when CONFIG_NO_BOOTMEM=y,  before
> > > > > > > e820__memblock_setup() is called there is no memory for bootmem
> > > > > > > to allocate,
> > > > > > 
> > > > > > Which you bloody well asked for by using NO_BOOTMEM=y.
> > > > > > 
> > > > > > Going down this route; adding hacks for every little thing that does
> > > > > > want bootmem, completely defeats the purpose.
> > > > > > 
> > > > > > If anything, make the earlycon thing depend on NO_BOOTMEM=n. That also
> > > > > > solves your problem. No earlycon, no panic.
> > > > > 
> > > > > Well, there is endeavor to remove bootmem allocator altogether. So
> > > > > making earlycon depend on NO_BOOTMEM=n doesn't sound like a good fit to
> > > > 
> > > > If we want to remove bootmem, then reintroducing it with a static bootmem
> > > > section doesn't make any sense at all.
> > > 
> > > I have actually checked the code now and see what you mean. I thought it
> > > would be a single allocation that is needed but that is not the case so
> > > the static buffer is not going to fly.
> > 
> > Exactly! I tried several ways for the static allocation, like in data, in bss
> > section, but they all failed, as in the very start of setup_arch():
> > 
> > 	memblock_reserve(__pa_symbol(_text),
> > 		(unsigned long)__bss_stop - (unsigned long)_text);
> > 
> > Those [_text, __bss_stop] is not able to be used by alloc_bootmem(). And I
> > only got this patch, and really appreciate any good suggestions.
> 
> And why do you want to use bootmem in the first place? If boot mem is going
> away, why are you not fixing the early console crap to NOT use bootmem at
> all?

The earlycon need use fixmap to map the mmio register, while fixmap will
need a new page for page table if the pmd/pte is not already their, as shown
in the commit log: 

	  panic+0xd2/0x220
	  __alloc_bootmem+0x31/0x34
	  spp_getpage+0x60/0x8a
	  fill_pte+0x71/0x130
	  __set_pte_vaddr+0x1d/0x50
	  set_pte_vaddr+0x3c/0x60
	  __native_set_fixmap+0x23/0x30
	  native_set_fixmap+0x30/0x40
	  setup_earlycon+0x1e0/0x32f
	  param_setup_earlycon+0x13/0x22
	  do_early_param+0x5b/0x90
	  parse_args+0x1f7/0x300
	  parse_early_options+0x24/0x28
	  parse_early_param+0x65/0x73
	  setup_arch+0x31e/0x9f1
	  start_kernel+0x58/0x44e

inside the spp_getpage():
	
	if (after_bootmem)
		ptr = (void *) get_zeroed_page(GFP_ATOMIC);
	else
		ptr = alloc_bootmem_pages(PAGE_SIZE);

Thanks,
Feng